convez.ai
MCP Server
leonard-cs

mcp-auth-bridge

leonard-cs/mcp-auth-bridge

Ver no GitHub
InicialPluginScore · de 0 a 100

A FastMCP proxy that wraps a header-auth-only remote MCP server, exposing it via OAuth or no-auth for clients that don't support header authentication.

Instalação

$npx convez-ai leonard-cs/mcp-auth-bridge

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

mcp-auth-bridge

A FastMCP proxy that wraps any MCP server — remote (header-auth or none) or local stdio — and re-exposes it over HTTP with header-auth, OAuth, or no-auth. It bridges servers to clients whose auth or transport support differs, and can fan several upstream servers into one endpoint.

MCP client ──HTTP (none | header | oauth)──▶  mcp-auth-bridge  ──▶  upstream MCP server(s)
                                                                     (stdio or remote, from .mcp.json)

Features

  • Three bridge auth modes: none, header (static bearer token), oauth.
  • OAuth, two ways: a local in-memory server for testing, or a generic remote verifier that works with any DCR-capable OIDC provider (WorkOS AuthKit, Descope, Keycloak, …).
  • Upstream from .mcp.json: the standard mcpServers schema (stdio or remote). Multiple servers are merged into one prefixed tool list.
  • Configured via .env — switch modes without touching code.

Quickstart

# 1. Install
uv sync

# 2. Configure the upstream server(s) you want to bridge
cp .mcp.json.example .mcp.json      # then edit: command/args/env or url/headers

# 3. Configure the bridge
cp .env.example .env                # then edit: AUTH_MODE, etc.

# 4. Run
uv run -m mcp_auth_bridge           # serves http://localhost:8000/mcp

Test it with the MCP Inspector:

npx @modelcontextprotocol/inspector
# Transport: Streamable HTTP   URL: http://localhost:8000/mcp

Configuration

All settings come from environment / .env (see .env.example):

VariableDefaultPurpose
AUTH_MODEnonenone | header | oauth
HOST0.0.0.0bind host
PORT8000bind port
MCP_CONFIG.mcp.jsonupstream definition (mcpServers schema)
OAUTH_PROVIDERmemorymemory (local test) | remote (real IdP)
BRIDGE_PUBLIC_URLhttp://localhost:8000public URL advertised in OAuth metadata
OAUTH_ISSUER—OIDC issuer (remote)
OAUTH_JWKS_URI—JWKS endpoint (remote)
OAUTH_AUTH_SERVER= issuerauthorization server to advertise
OAUTH_AUDIENCE—expected token audience (blank = skip)
OAUTH_ALGORITHMRS256JWT signing algorithm

Auth modes

none

No authentication — the bridge is an open proxy. Useful for local development.

header

Clients must send Authorization: Bearer <token>. The accepted token(s) are defined in src/mcp_auth_bridge/auth/header.py (BRIDGE_TOKENS). Replace the sample token before any real use.

oauth

Full OAuth 2.1. Selected sub-mode via OAUTH_PROVIDER:

  • memory — the bridge is its own in-memory OAuth server (DCR, authorize, token exchange, auto-approved). No external account; works on localhost. Testing/dev only.
  • remote — the bridge verifies JWTs from a real external IdP and advertises it to clients. Provider-agnostic; configure via OAUTH_*.

To set up oauth + remote with WorkOS AuthKit (incl. the ngrok tunnel and how to register / refresh the public URL), see docs/authkit-ngrok-setup.md.

Project layout

src/mcp_auth_bridge/
  __main__.py      # entrypoint: reads host/port, runs the proxy
  server.py        # build_server(): load upstream + auth, create proxy
  config.py        # Settings (pydantic-settings), loads .env
  upstream.py      # load_upstream_config(): reads .mcp.json
  auth/
    __init__.py    # build_auth_provider(settings): dispatch on AUTH_MODE
    none.py        # no auth
    header.py      # static bearer token (StaticTokenVerifier)
    oauth.py       # in-memory + generic remote OAuth
Ver o original no GitHub

Bom para

Relacionados